Infrastructure

Everything around the GPU matters as much as the GPU itself once you're running production infrastructure instead of a single instance: how jobs get scheduled, how nodes talk to each other, how storage keeps up with checkpoint I/O, and how you know something is wrong before a customer notices. This cluster covers that layer.

On scheduling and orchestration: Kubernetes GPU orchestration with DRA and the KAI Scheduler, Slurm for HPC-style batch jobs, SkyPilot for multi-cloud cost-aware placement, and NVIDIA's newer Grove and Mission Control stacks for disaggregated inference. On networking: NVLink versus PCIe versus the newer UALink standard, and InfiniBand versus RoCE versus Spectrum-X for training fabrics. On storage: WekaIO, Lustre, and BeeGFS for multi-node training, and GPU Direct Storage for checkpoint and weight loading that skips the CPU bottleneck.

We also cover observability directly. The pillar post, GPU Monitoring for ML, walks through nvidia-smi, DCGM with Prometheus, and Grafana dashboards for clusters where most teams are running under 70% utilization without knowing it. Start there if you don't yet have visibility into what your GPUs are doing, then move into the scheduling or networking posts that match your bottleneck.